Musubi unveils PolicyLM-1.7B for real-time content moderation

Content moderation has a speed problem. Posts go live in an instant, and the systems meant to police them often lag behind.

Musubi is betting a smaller, faster model can close that gap. On Tuesday, the trust and safety startup announced PolicyLM-1.7B, a lightweight decision model built for real-time moderation and released with open weights.

What Musubi actually shipped

PolicyLM-1.7B is a classifier, not a chatbot. Its job is to look at a piece of content, check it against a platform’s rules, and return a verdict.

The key feature is that those rules are custom. Platforms can feed the model their own policies instead of relying on a one-size-fits-all definition of harmful content.

According to Musubi’s press materials, PolicyLM can classify content in under 100 milliseconds.

The model targets high-volume platforms where latency and infrastructure costs are central concerns.

Open weights mean developers can download the trained model, inspect it and run it on their own hardware rather than renting access through an API.

Part of a bigger toolkit

PolicyLM does not stand alone. It slots into a broader Musubi suite that covers several layers of the trust and safety stack.

That lineup includes PolicyAI, which handles moderation driven by large language models. It also includes AiMod and FraudAI for fraud detection, plus a Media Forensics tool.

Then there is Musubi Coop, which provides a human review workflow.

Musubi says its technology achieves F1 scores exceeding 90% on classification tasks. F1 is a metric that weighs two things at once: how much bad content a system catches, and how often it wrongly flags innocent posts.

The company also reports a striking operational payoff. Clients including Bluesky and Grindr have reportedly seen reductions of 70-95% in manual review requirements.

Who is behind Musubi

Musubi was founded in 2023 by Tom Quisel and Filip Jankovic. Since launching, the company has raised approximately $8 million.

Bluesky and Grindr both run platforms where moderation is not optional, and Musubi says its tools protect hundreds of millions of users.

In March 2026, Musubi integrated NVIDIA NeMo Guardrails, a framework for keeping AI systems within defined boundaries.

Later that year it introduced the Musubi Agent, which automates policy iteration.

What this means for platforms and the moderation market

The open-weights approach shifts the competitive dynamics. By giving the model away, Musubi lowers the barrier for platforms to try its technology, then can sell them the surrounding tools and services.

Open weights mean anyone can study how the model makes decisions, including bad actors looking for ways to slip content past it.

Custom policies also put the burden of clarity on platforms. A fast model enforcing a poorly written rule will simply make mistakes faster.

Musubi’s policy-driven design means swapping in an updated rule is easier than retraining a model from scratch.